A security researcher who goes by bobdahacker has published details of a vulnerability in tl;dv, an AI meeting recording platform with over 2 million users, that left 181,874 meeting records accessible to any authenticated user on the platform. The vulnerability was disclosed to tl;dv in January 2026. Six months later, it was still unpatched.
The exposure covered meeting records from 84,312 users across 35,003 email domains. Government agencies in 23 countries — including Brazil, Ukraine, the Philippines, Malaysia, Japan, and the United States — had their meeting metadata readable by any free-tier tl;dv user. Universities including Berkeley and the University of Tokyo were in the database. So were corporations like HubSpot, Confluent, and Mitsui Fudosan.
What is tl;dv? It’s an AI meeting recorder. You drop a bot into a Google Meet, Zoom, or Teams call, and it records everything, transcribes it, and generates AI summaries. Sales calls, job interviews, performance reviews, internal strategy sessions — the kind of meetings where people share trade secrets for 45 minutes after someone nervously says “this call is being recorded.”
How It Worked
The vulnerability was a missing tenant isolation control in tl;dv’s Firestore database. When a user signed up, they received a Firebase token that let them query the meetings collection. There was no check limiting queries to the user’s own meetings. Any authenticated user could query every meeting across every account on the platform.
Each meeting record exposed the creator’s email address, the conference ID (a joinable Google Meet or Teams room), the provider, recording status, and timestamps. For meetings in recording status — roughly 1,000 at any given time — that conference ID was a live, active call. A user could watch the collection in real time, see a meeting start recording, grab the ID, and walk into someone else’s call uninvited.
The researcher demonstrated this by joining two live meetings: a Malaysian Ministry of Education presentation with 157 participants, and a university startup project call with 21 people screen-sharing their work. In both cases, nobody invited the researcher. The Firestore database did.
The Compliance Contradiction
What makes this story more than a standard data breach is that tl;dv publicly claimed compliance with GDPR, SOC2, and the EU AI Act. The AI Governance Institute notes that these compliance claims “directly contradict” the sustained exposure — a vendor can simultaneously publish SOC2 and GDPR compliance assertions while maintaining a six-month-old critical data exposure.
This is the part that should make any organisation using AI-powered productivity tools uncomfortable. The procurement-stage questionnaires and self-attestations that most enterprises rely on to vet vendors failed entirely as a verification mechanism. tl;dv ticked the compliance boxes. The vulnerability sat open for half a year.
For organisations in GDPR-covered jurisdictions, the six-month remediation gap likely triggers mandatory breach notification obligations — and not just for tl;dv. Affected enterprises that are deemed data controllers over meeting content may need to assess whether they need to notify their own supervisory authorities under GDPR Articles 33 and 34.
The Broader Pattern
AI meeting recorders occupy a uniquely risky position in the data supply chain. They capture unfiltered executive conversations, legal discussions, HR matters, and customer calls. Most organisations classify them as low-stakes productivity tools and apply minimal ongoing oversight after the initial procurement approval.
The tl;dv incident is not isolated. It reflects a broader pattern in multi-tenant SaaS architectures that handle unstructured personal data at scale. The same class of vulnerability — missing tenant isolation in a shared database — has been at the root of numerous SaaS breaches across both AI and non-AI platforms. What’s different here is the combination: AI-grade sensitivity of the captured content (transcripts, summaries, real-time audio) plus the compliance claims that masked the problem.
The EU’s Article 50 transparency rules, which took effect on August 2, require AI systems to disclose their AI nature. But transparency about what a system is doesn’t address whether the system works. That’s a different regulatory question, and one that the EU AI Office is likely to scrutinise more closely as incidents like this accumulate.
What Stands Out
The researcher’s blog post is worth reading for the detail alone — including the discovery that tl;dv names its microservices after pasta (cappellini, carbonara, fusilli, penne, puttanesca, ravioli) and that an internal FIFA World Cup prediction game had its own unauthenticated API leaking employee names and emails. But the serious finding is the gap between what the company claimed and what the platform actually delivered.
A compliance badge is not a control. A SOC2 attestation is a point-in-time snapshot, not a continuous guarantee. And when a vendor that claims EU AI Act compliance can sustain a critical vulnerability for six months without remediation, the question isn’t whether tl;dv has a security problem. It’s whether the entire framework of vendor self-attestation is adequate for tools that ingest unfiltered business conversations at scale.
For New Zealand organisations using AI meeting tools — and there are many, given the prevalence of remote work and cross-border calls — the lesson is direct. Audit what’s deployed, including tools adopted informally by employees. Check whether the vendor has been independently verified, not just self-attested. And treat AI meeting recorders as high-risk data intake tools, not productivity widgets.
📰 Sources
- bobdahacker — tl;dv (Too Lazy; Didn’t Validate): 181,874 Meetings Left Wide Open
- AI Governance Institute — 181,874 Meetings Exposed After tl;dv Ignored Six-Month Disclosure
- The Register — Gym rat asks AI agent to book him a class, it hacks a waitlist API
- explainx.ai — tl;dv Firestore Breach: 181,874 Meetings Exposed
